Hugging Face and AWS partner to make AI open and accessible

Last week, Adam Selipsky shared that AWS is expanding its partnership with Hugging Face to make AI open, accessible, and affordable to enable customers to create high-performance, low-cost generative AI apps. Learn more about it in the announcement blog post.

News & Announcements 📣

MetaAI introduced LLaMA, a foundational, 65-billion-parameter large language model with GPT-3 like performance and a license that forbids any commercial use.

Sasha Rush open-sourced an LLM chaining & prompting library called Mini-Chain, similar to LangChain, with the goal of implementing the core prompt chaining functionality in a tiny digestible library.

IntelAI open-sourced a new fastRAG, a new library to build retrieval-augmented generative models and applications e.g., for question answering and summarization.

Microsoft released E5, a model family for text embeddings to build search applications achieving top performances on MTEB and BEIR benchmarks.

Cohere introduced its new summarization endpoint, which allows users to summarize documents with up to 50,000 characters.

Timm (Pytorch image models) has officially joined the Hugging Face Libraries family.

Tutorials & Demos 📝

Hundredblocks created a practical example of how model parallelism works, exploring data parallelism, tensor parallelism, and pipeline parallelism.

I created an example of how to fine-tune FLAN-T5 XL & XXL using Deepspeed and Hugging Face Transformers with an extended version on how to use Amazon SageMaker as infrastructure and combine it with DeepSpeed.

Toshihiro Hayashi created an unofficial demo for ControlNet for controlled text-to-image generation with Stable Diffusion.

Sylvain Filoni created a Pix2Pix Video demo to apply the Pix2Pix diffusion to videos. Check out the space for some awesome examples.

Sourab added an example to PEFT on how to efficiently fine-tune Whisper-large using PEFT.

Yannic Klicher did some hands-on coding for adding token-streaming to text-generation-inference and Open Assistant.

Julien Simon recorded an awesome video on training Transformers with AWS Trainium and the Hugging Face Neuron AMI.

Dair AI continues to improve its Prompt Engineering Guide, including guides, tools, blogs, and papers.

Reads & Papers 📚

Deepset wrote and nice blog post on when and how to train Language Models.

OpenAI shared their thoughts and plans for AGI and beyond.

Aligning Text-to-Image Models using Human Feedback papers got published showing how to improve image generation by human feedback.

GitHub shares new insights on Copilot and how much better it has become since the initial release.

Composer: Creative and Controllable Image Synthesis with Composable Conditions introduces a 5 billion parameter controllable diffusion model for more controllable image generation.

I hope you enjoyed this newsletter. 🤗 If you have any questions or are interested in collaborating, feel free to contact me on Twitter or LinkedIn.

See you next week 👋🏻👋🏻